A Multi-Feature Fusion Approach for ECG Signal Classification

Arrhythmia is a typical condition in cardiovascular diseases, posing a serious threat to patients' health. Electrocardiogram (ECG), as an effective tool for detecting arrhythmia, plays a crucial role in diagnosing the condition through automatic classification. However, most existing ECG classification methods focus on extracting features from time-space domain signals, often overlooking the significant value of frequency-domain information. This study proposes a novel ECG signal analysis method aimed at integrating time, space, and frequency domain features. Specifically, we first utilize multi-scale dilated convolutional modules to capture and fuse spatial features at different scales. Next, we introduce a frequency-domain perception module to extract the frequency-domain characteristics of ECG signals. To further enhance the interaction between time-space features, a bidirectional inversion mechanism is introduced to generate forward and backward sequences for both spatial and frequency domain features. These are then deeply fused through a bidirectional self-attention mechanism, enabling comprehensive extraction of contextual information. Experimental results show that, the proposed feature fusion strategy significantly improves the recognition ability of ECG signals, demonstrating high robustness and potential for application.